基于全卷积神经网络的图像去雾算法

针对雾天环境下采集的图像对比度降低,饱和度下降以及色彩偏移问题,提出了一种基于全卷积神经网络的图像去雾算法。首先,提出的三个尺度的全卷积神经网络用来学习雾天图像与介质传输图之间的映射关系,逐步生成精细的介质传输图;其次,通过雾天图像引导滤波优化预测的介质传输图,使得图像边缘信息更加平滑;最后,根据暗原色先验理论估计大气光的值,通过大气散射模型恢复出无雾图像。该方法获得的无雾图像不但未造成图像中有用信息的损失,并且恢复的图像色彩自然。实验结果表明,该去雾算法在自然雾天图像和利用Middlebury Stereo Datasets合成的雾天图像上均优于其他对比算法,恢复的图像具有更好的对比度与清晰度。

Image defogging algorithm combined with full convolution neural network

Aiming at the problems of contrast reduction, saturation reduction and color migration of images collected in foggy environment, an image defogging algorithm based on full convolution neural network is put forward. First, the proposed three scales convolution neural network is used to study the fog of the mapping relationship between foggy image and medium transmission map, gradually produce the refine medium transmission map; secondly, the foggy image is used as a guide map to refine the forecasting medium transmission map, so as to make the edge information of the image more smooth; finally, the value of atmospheric light is estimated according to the dark channel prior theory, and the fog-free image is recovered by the atmospheric scattering model. The fog-free image obtained by this method not only causes no loss of useful information in the image, but also restores the color of the image naturally. Experimental results show that the algorithm proposed is superior to other comparison algorithms in both natural fog images and fog images produced by Middlebury Stereo Datasets, and the restored images have better contrast and clarity.